Master Thesis

“Large Language Models for Procedural Animations in Virtual Environments”

The goal of this thesis is to investigate how large language models (LLMs) can be utilized to generate context-aware procedural behaviors for avatars in human-centric simulation environments. In simulation environments where robots interact with avatars and virtual agents, LLMs can serve as high-level reasoning modules that translate natural-language context into structured action plans. These plans govern how virtual agents behave, interact, and adapt to dynamic environments, supporting workflows such as task coordination, object manipulation, or situational response.

Traditional systems rely on manually scripted rules or stochastic animation graphs, which are often rigid and difficult to scale. This thesis explores whether LLM-guided high-level planning can provide a more flexible, semantically grounded, and scalable alternative. Specifically, the work aims to:

• Develop methods for using LLMs to produce symbolic behavior plans that describe what an avatar should do and how (e.g., action type, speed, style, timing).
• Integrate these symbolic plans with a motion execution layer in Unity that ensures physical fidelity, safety, and reproducibility.
• Evaluate the feasibility, benefits, and limitations of this LLM-driven procedural control approach across a small set of representative simulation scenarios.

The expected outcome is a validated prototype pipeline demonstrating how LLM-based behavior generation can improve semantic understanding, adaptability, and autonomy in simulation-based environments.